vm_pageout.c revision 207374
1/*- 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 1994 John S. Dyson 5 * All rights reserved. 6 * Copyright (c) 1994 David Greenman 7 * All rights reserved. 8 * Copyright (c) 2005 Yahoo! Technologies Norway AS 9 * All rights reserved. 10 * 11 * This code is derived from software contributed to Berkeley by 12 * The Mach Operating System project at Carnegie-Mellon University. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 43 * 44 * 45 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 46 * All rights reserved. 47 * 48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 49 * 50 * Permission to use, copy, modify and distribute this software and 51 * its documentation is hereby granted, provided that both the copyright 52 * notice and this permission notice appear in all copies of the 53 * software, derivative works or modified versions, and any portions 54 * thereof, and that both notices appear in supporting documentation. 55 * 56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 57 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 59 * 60 * Carnegie Mellon requests users of this software to return to 61 * 62 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 63 * School of Computer Science 64 * Carnegie Mellon University 65 * Pittsburgh PA 15213-3890 66 * 67 * any improvements or extensions that they make and grant Carnegie the 68 * rights to redistribute these changes. 69 */ 70 71/* 72 * The proverbial page-out daemon. 73 */ 74 75#include <sys/cdefs.h> 76__FBSDID("$FreeBSD: head/sys/vm/vm_pageout.c 207374 2010-04-29 16:18:45Z alc $"); 77 78#include "opt_vm.h" 79#include <sys/param.h> 80#include <sys/systm.h> 81#include <sys/kernel.h> 82#include <sys/eventhandler.h> 83#include <sys/lock.h> 84#include <sys/mutex.h> 85#include <sys/proc.h> 86#include <sys/kthread.h> 87#include <sys/ktr.h> 88#include <sys/mount.h> 89#include <sys/resourcevar.h> 90#include <sys/sched.h> 91#include <sys/signalvar.h> 92#include <sys/vnode.h> 93#include <sys/vmmeter.h> 94#include <sys/sx.h> 95#include <sys/sysctl.h> 96 97#include <vm/vm.h> 98#include <vm/vm_param.h> 99#include <vm/vm_object.h> 100#include <vm/vm_page.h> 101#include <vm/vm_map.h> 102#include <vm/vm_pageout.h> 103#include <vm/vm_pager.h> 104#include <vm/swap_pager.h> 105#include <vm/vm_extern.h> 106#include <vm/uma.h> 107 108/* 109 * System initialization 110 */ 111 112/* the kernel process "vm_pageout"*/ 113static void vm_pageout(void); 114static int vm_pageout_clean(vm_page_t); 115static void vm_pageout_scan(int pass); 116 117struct proc *pageproc; 118 119static struct kproc_desc page_kp = { 120 "pagedaemon", 121 vm_pageout, 122 &pageproc 123}; 124SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, 125 &page_kp); 126 127#if !defined(NO_SWAPPING) 128/* the kernel process "vm_daemon"*/ 129static void vm_daemon(void); 130static struct proc *vmproc; 131 132static struct kproc_desc vm_kp = { 133 "vmdaemon", 134 vm_daemon, 135 &vmproc 136}; 137SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); 138#endif 139 140 141int vm_pages_needed; /* Event on which pageout daemon sleeps */ 142int vm_pageout_deficit; /* Estimated number of pages deficit */ 143int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 144 145#if !defined(NO_SWAPPING) 146static int vm_pageout_req_swapout; /* XXX */ 147static int vm_daemon_needed; 148static struct mtx vm_daemon_mtx; 149/* Allow for use by vm_pageout before vm_daemon is initialized. */ 150MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); 151#endif 152static int vm_max_launder = 32; 153static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 154static int vm_pageout_full_stats_interval = 0; 155static int vm_pageout_algorithm=0; 156static int defer_swap_pageouts=0; 157static int disable_swap_pageouts=0; 158 159#if defined(NO_SWAPPING) 160static int vm_swap_enabled=0; 161static int vm_swap_idle_enabled=0; 162#else 163static int vm_swap_enabled=1; 164static int vm_swap_idle_enabled=0; 165#endif 166 167SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 168 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 169 170SYSCTL_INT(_vm, OID_AUTO, max_launder, 171 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 172 173SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 174 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 175 176SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 177 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 178 179SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 180 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 181 182#if defined(NO_SWAPPING) 183SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 184 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout"); 185SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 186 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 187#else 188SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 189 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 190SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 191 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 192#endif 193 194SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 195 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 196 197SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 198 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 199 200static int pageout_lock_miss; 201SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 202 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 203 204#define VM_PAGEOUT_PAGE_COUNT 16 205int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 206 207int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 208SYSCTL_INT(_vm, OID_AUTO, max_wired, 209 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); 210 211#if !defined(NO_SWAPPING) 212static void vm_pageout_map_deactivate_pages(vm_map_t, long); 213static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 214static void vm_req_vmdaemon(int req); 215#endif 216static void vm_pageout_page_stats(void); 217 218/* 219 * vm_pageout_fallback_object_lock: 220 * 221 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is 222 * known to have failed and page queue must be either PQ_ACTIVE or 223 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 224 * while locking the vm object. Use marker page to detect page queue 225 * changes and maintain notion of next page on page queue. Return 226 * TRUE if no changes were detected, FALSE otherwise. vm object is 227 * locked on return. 228 * 229 * This function depends on both the lock portion of struct vm_object 230 * and normal struct vm_page being type stable. 231 */ 232boolean_t 233vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 234{ 235 struct vm_page marker; 236 boolean_t unchanged; 237 u_short queue; 238 vm_object_t object; 239 240 /* 241 * Initialize our marker 242 */ 243 bzero(&marker, sizeof(marker)); 244 marker.flags = PG_FICTITIOUS | PG_MARKER; 245 marker.oflags = VPO_BUSY; 246 marker.queue = m->queue; 247 marker.wire_count = 1; 248 249 queue = m->queue; 250 object = m->object; 251 252 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, 253 m, &marker, pageq); 254 vm_page_unlock_queues(); 255 VM_OBJECT_LOCK(object); 256 vm_page_lock_queues(); 257 258 /* Page queue might have changed. */ 259 *next = TAILQ_NEXT(&marker, pageq); 260 unchanged = (m->queue == queue && 261 m->object == object && 262 &marker == TAILQ_NEXT(m, pageq)); 263 TAILQ_REMOVE(&vm_page_queues[queue].pl, 264 &marker, pageq); 265 return (unchanged); 266} 267 268/* 269 * vm_pageout_clean: 270 * 271 * Clean the page and remove it from the laundry. 272 * 273 * We set the busy bit to cause potential page faults on this page to 274 * block. Note the careful timing, however, the busy bit isn't set till 275 * late and we cannot do anything that will mess with the page. 276 */ 277static int 278vm_pageout_clean(m) 279 vm_page_t m; 280{ 281 vm_object_t object; 282 vm_page_t mc[2*vm_pageout_page_count]; 283 int pageout_count; 284 int ib, is, page_base; 285 vm_pindex_t pindex = m->pindex; 286 287 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 288 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 289 290 /* 291 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 292 * with the new swapper, but we could have serious problems paging 293 * out other object types if there is insufficient memory. 294 * 295 * Unfortunately, checking free memory here is far too late, so the 296 * check has been moved up a procedural level. 297 */ 298 299 /* 300 * Can't clean the page if it's busy or held. 301 */ 302 if ((m->hold_count != 0) || 303 ((m->busy != 0) || (m->oflags & VPO_BUSY))) { 304 return 0; 305 } 306 307 mc[vm_pageout_page_count] = m; 308 pageout_count = 1; 309 page_base = vm_pageout_page_count; 310 ib = 1; 311 is = 1; 312 313 /* 314 * Scan object for clusterable pages. 315 * 316 * We can cluster ONLY if: ->> the page is NOT 317 * clean, wired, busy, held, or mapped into a 318 * buffer, and one of the following: 319 * 1) The page is inactive, or a seldom used 320 * active page. 321 * -or- 322 * 2) we force the issue. 323 * 324 * During heavy mmap/modification loads the pageout 325 * daemon can really fragment the underlying file 326 * due to flushing pages out of order and not trying 327 * align the clusters (which leave sporatic out-of-order 328 * holes). To solve this problem we do the reverse scan 329 * first and attempt to align our cluster, then do a 330 * forward scan if room remains. 331 */ 332 object = m->object; 333more: 334 while (ib && pageout_count < vm_pageout_page_count) { 335 vm_page_t p; 336 337 if (ib > pindex) { 338 ib = 0; 339 break; 340 } 341 342 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 343 ib = 0; 344 break; 345 } 346 if ((p->oflags & VPO_BUSY) || p->busy) { 347 ib = 0; 348 break; 349 } 350 vm_page_test_dirty(p); 351 if (p->dirty == 0 || 352 p->queue != PQ_INACTIVE || 353 p->hold_count != 0) { /* may be undergoing I/O */ 354 ib = 0; 355 break; 356 } 357 mc[--page_base] = p; 358 ++pageout_count; 359 ++ib; 360 /* 361 * alignment boundry, stop here and switch directions. Do 362 * not clear ib. 363 */ 364 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 365 break; 366 } 367 368 while (pageout_count < vm_pageout_page_count && 369 pindex + is < object->size) { 370 vm_page_t p; 371 372 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 373 break; 374 if ((p->oflags & VPO_BUSY) || p->busy) { 375 break; 376 } 377 vm_page_test_dirty(p); 378 if (p->dirty == 0 || 379 p->queue != PQ_INACTIVE || 380 p->hold_count != 0) { /* may be undergoing I/O */ 381 break; 382 } 383 mc[page_base + pageout_count] = p; 384 ++pageout_count; 385 ++is; 386 } 387 388 /* 389 * If we exhausted our forward scan, continue with the reverse scan 390 * when possible, even past a page boundry. This catches boundry 391 * conditions. 392 */ 393 if (ib && pageout_count < vm_pageout_page_count) 394 goto more; 395 396 /* 397 * we allow reads during pageouts... 398 */ 399 return (vm_pageout_flush(&mc[page_base], pageout_count, 0)); 400} 401 402/* 403 * vm_pageout_flush() - launder the given pages 404 * 405 * The given pages are laundered. Note that we setup for the start of 406 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 407 * reference count all in here rather then in the parent. If we want 408 * the parent to do more sophisticated things we may have to change 409 * the ordering. 410 */ 411int 412vm_pageout_flush(vm_page_t *mc, int count, int flags) 413{ 414 vm_object_t object = mc[0]->object; 415 int pageout_status[count]; 416 int numpagedout = 0; 417 int i; 418 419 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 420 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 421 /* 422 * Initiate I/O. Bump the vm_page_t->busy counter and 423 * mark the pages read-only. 424 * 425 * We do not have to fixup the clean/dirty bits here... we can 426 * allow the pager to do it after the I/O completes. 427 * 428 * NOTE! mc[i]->dirty may be partial or fragmented due to an 429 * edge case with file fragments. 430 */ 431 for (i = 0; i < count; i++) { 432 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 433 ("vm_pageout_flush: partially invalid page %p index %d/%d", 434 mc[i], i, count)); 435 vm_page_io_start(mc[i]); 436 pmap_remove_write(mc[i]); 437 } 438 vm_page_unlock_queues(); 439 vm_object_pip_add(object, count); 440 441 vm_pager_put_pages(object, mc, count, flags, pageout_status); 442 443 vm_page_lock_queues(); 444 for (i = 0; i < count; i++) { 445 vm_page_t mt = mc[i]; 446 447 KASSERT(pageout_status[i] == VM_PAGER_PEND || 448 (mt->flags & PG_WRITEABLE) == 0, 449 ("vm_pageout_flush: page %p is not write protected", mt)); 450 switch (pageout_status[i]) { 451 case VM_PAGER_OK: 452 case VM_PAGER_PEND: 453 numpagedout++; 454 break; 455 case VM_PAGER_BAD: 456 /* 457 * Page outside of range of object. Right now we 458 * essentially lose the changes by pretending it 459 * worked. 460 */ 461 vm_page_undirty(mt); 462 break; 463 case VM_PAGER_ERROR: 464 case VM_PAGER_FAIL: 465 /* 466 * If page couldn't be paged out, then reactivate the 467 * page so it doesn't clog the inactive list. (We 468 * will try paging out it again later). 469 */ 470 vm_page_activate(mt); 471 break; 472 case VM_PAGER_AGAIN: 473 break; 474 } 475 476 /* 477 * If the operation is still going, leave the page busy to 478 * block all other accesses. Also, leave the paging in 479 * progress indicator set so that we don't attempt an object 480 * collapse. 481 */ 482 if (pageout_status[i] != VM_PAGER_PEND) { 483 vm_object_pip_wakeup(object); 484 vm_page_io_finish(mt); 485 if (vm_page_count_severe()) 486 vm_page_try_to_cache(mt); 487 } 488 } 489 return numpagedout; 490} 491 492#if !defined(NO_SWAPPING) 493/* 494 * vm_pageout_object_deactivate_pages 495 * 496 * deactivate enough pages to satisfy the inactive target 497 * requirements or if vm_page_proc_limit is set, then 498 * deactivate all of the pages in the object and its 499 * backing_objects. 500 * 501 * The object and map must be locked. 502 */ 503static void 504vm_pageout_object_deactivate_pages(pmap, first_object, desired) 505 pmap_t pmap; 506 vm_object_t first_object; 507 long desired; 508{ 509 vm_object_t backing_object, object; 510 vm_page_t p, next; 511 int actcount, remove_mode; 512 513 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 514 if (first_object->type == OBJT_DEVICE || 515 first_object->type == OBJT_SG) 516 return; 517 for (object = first_object;; object = backing_object) { 518 if (pmap_resident_count(pmap) <= desired) 519 goto unlock_return; 520 if (object->type == OBJT_PHYS || object->paging_in_progress) 521 goto unlock_return; 522 523 remove_mode = 0; 524 if (object->shadow_count > 1) 525 remove_mode = 1; 526 /* 527 * scan the objects entire memory queue 528 */ 529 p = TAILQ_FIRST(&object->memq); 530 vm_page_lock_queues(); 531 while (p != NULL) { 532 if (pmap_resident_count(pmap) <= desired) { 533 vm_page_unlock_queues(); 534 goto unlock_return; 535 } 536 next = TAILQ_NEXT(p, listq); 537 cnt.v_pdpages++; 538 if (p->wire_count != 0 || 539 p->hold_count != 0 || 540 p->busy != 0 || 541 (p->oflags & VPO_BUSY) || 542 !pmap_page_exists_quick(pmap, p)) { 543 p = next; 544 continue; 545 } 546 actcount = pmap_ts_referenced(p); 547 if (actcount) { 548 vm_page_flag_set(p, PG_REFERENCED); 549 } else if (p->flags & PG_REFERENCED) { 550 actcount = 1; 551 } 552 if ((p->queue != PQ_ACTIVE) && 553 (p->flags & PG_REFERENCED)) { 554 vm_page_activate(p); 555 p->act_count += actcount; 556 vm_page_flag_clear(p, PG_REFERENCED); 557 } else if (p->queue == PQ_ACTIVE) { 558 if ((p->flags & PG_REFERENCED) == 0) { 559 p->act_count -= min(p->act_count, ACT_DECLINE); 560 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 561 pmap_remove_all(p); 562 vm_page_deactivate(p); 563 } else { 564 vm_page_requeue(p); 565 } 566 } else { 567 vm_page_activate(p); 568 vm_page_flag_clear(p, PG_REFERENCED); 569 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 570 p->act_count += ACT_ADVANCE; 571 vm_page_requeue(p); 572 } 573 } else if (p->queue == PQ_INACTIVE) { 574 pmap_remove_all(p); 575 } 576 p = next; 577 } 578 vm_page_unlock_queues(); 579 if ((backing_object = object->backing_object) == NULL) 580 goto unlock_return; 581 VM_OBJECT_LOCK(backing_object); 582 if (object != first_object) 583 VM_OBJECT_UNLOCK(object); 584 } 585unlock_return: 586 if (object != first_object) 587 VM_OBJECT_UNLOCK(object); 588} 589 590/* 591 * deactivate some number of pages in a map, try to do it fairly, but 592 * that is really hard to do. 593 */ 594static void 595vm_pageout_map_deactivate_pages(map, desired) 596 vm_map_t map; 597 long desired; 598{ 599 vm_map_entry_t tmpe; 600 vm_object_t obj, bigobj; 601 int nothingwired; 602 603 if (!vm_map_trylock(map)) 604 return; 605 606 bigobj = NULL; 607 nothingwired = TRUE; 608 609 /* 610 * first, search out the biggest object, and try to free pages from 611 * that. 612 */ 613 tmpe = map->header.next; 614 while (tmpe != &map->header) { 615 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 616 obj = tmpe->object.vm_object; 617 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 618 if (obj->shadow_count <= 1 && 619 (bigobj == NULL || 620 bigobj->resident_page_count < obj->resident_page_count)) { 621 if (bigobj != NULL) 622 VM_OBJECT_UNLOCK(bigobj); 623 bigobj = obj; 624 } else 625 VM_OBJECT_UNLOCK(obj); 626 } 627 } 628 if (tmpe->wired_count > 0) 629 nothingwired = FALSE; 630 tmpe = tmpe->next; 631 } 632 633 if (bigobj != NULL) { 634 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 635 VM_OBJECT_UNLOCK(bigobj); 636 } 637 /* 638 * Next, hunt around for other pages to deactivate. We actually 639 * do this search sort of wrong -- .text first is not the best idea. 640 */ 641 tmpe = map->header.next; 642 while (tmpe != &map->header) { 643 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 644 break; 645 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 646 obj = tmpe->object.vm_object; 647 if (obj != NULL) { 648 VM_OBJECT_LOCK(obj); 649 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 650 VM_OBJECT_UNLOCK(obj); 651 } 652 } 653 tmpe = tmpe->next; 654 } 655 656 /* 657 * Remove all mappings if a process is swapped out, this will free page 658 * table pages. 659 */ 660 if (desired == 0 && nothingwired) { 661 pmap_remove(vm_map_pmap(map), vm_map_min(map), 662 vm_map_max(map)); 663 } 664 vm_map_unlock(map); 665} 666#endif /* !defined(NO_SWAPPING) */ 667 668/* 669 * vm_pageout_scan does the dirty work for the pageout daemon. 670 */ 671static void 672vm_pageout_scan(int pass) 673{ 674 vm_page_t m, next; 675 struct vm_page marker; 676 int page_shortage, maxscan, pcount; 677 int addl_page_shortage, addl_page_shortage_init; 678 vm_object_t object; 679 int actcount; 680 int vnodes_skipped = 0; 681 int maxlaunder; 682 683 /* 684 * Decrease registered cache sizes. 685 */ 686 EVENTHANDLER_INVOKE(vm_lowmem, 0); 687 /* 688 * We do this explicitly after the caches have been drained above. 689 */ 690 uma_reclaim(); 691 692 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 693 694 /* 695 * Calculate the number of pages we want to either free or move 696 * to the cache. 697 */ 698 page_shortage = vm_paging_target() + addl_page_shortage_init; 699 700 /* 701 * Initialize our marker 702 */ 703 bzero(&marker, sizeof(marker)); 704 marker.flags = PG_FICTITIOUS | PG_MARKER; 705 marker.oflags = VPO_BUSY; 706 marker.queue = PQ_INACTIVE; 707 marker.wire_count = 1; 708 709 /* 710 * Start scanning the inactive queue for pages we can move to the 711 * cache or free. The scan will stop when the target is reached or 712 * we have scanned the entire inactive queue. Note that m->act_count 713 * is not used to form decisions for the inactive queue, only for the 714 * active queue. 715 * 716 * maxlaunder limits the number of dirty pages we flush per scan. 717 * For most systems a smaller value (16 or 32) is more robust under 718 * extreme memory and disk pressure because any unnecessary writes 719 * to disk can result in extreme performance degredation. However, 720 * systems with excessive dirty pages (especially when MAP_NOSYNC is 721 * used) will die horribly with limited laundering. If the pageout 722 * daemon cannot clean enough pages in the first pass, we let it go 723 * all out in succeeding passes. 724 */ 725 if ((maxlaunder = vm_max_launder) <= 1) 726 maxlaunder = 1; 727 if (pass) 728 maxlaunder = 10000; 729 vm_page_lock_queues(); 730rescan0: 731 addl_page_shortage = addl_page_shortage_init; 732 maxscan = cnt.v_inactive_count; 733 734 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 735 m != NULL && maxscan-- > 0 && page_shortage > 0; 736 m = next) { 737 738 cnt.v_pdpages++; 739 740 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) { 741 goto rescan0; 742 } 743 744 next = TAILQ_NEXT(m, pageq); 745 object = m->object; 746 747 /* 748 * skip marker pages 749 */ 750 if (m->flags & PG_MARKER) 751 continue; 752 753 /* 754 * A held page may be undergoing I/O, so skip it. 755 */ 756 if (m->hold_count) { 757 vm_page_requeue(m); 758 addl_page_shortage++; 759 continue; 760 } 761 /* 762 * Don't mess with busy pages, keep in the front of the 763 * queue, most likely are being paged out. 764 */ 765 if (!VM_OBJECT_TRYLOCK(object) && 766 (!vm_pageout_fallback_object_lock(m, &next) || 767 m->hold_count != 0)) { 768 VM_OBJECT_UNLOCK(object); 769 addl_page_shortage++; 770 continue; 771 } 772 if (m->busy || (m->oflags & VPO_BUSY)) { 773 VM_OBJECT_UNLOCK(object); 774 addl_page_shortage++; 775 continue; 776 } 777 778 /* 779 * If the object is not being used, we ignore previous 780 * references. 781 */ 782 if (object->ref_count == 0) { 783 vm_page_flag_clear(m, PG_REFERENCED); 784 KASSERT(!pmap_page_is_mapped(m), 785 ("vm_pageout_scan: page %p is mapped", m)); 786 787 /* 788 * Otherwise, if the page has been referenced while in the 789 * inactive queue, we bump the "activation count" upwards, 790 * making it less likely that the page will be added back to 791 * the inactive queue prematurely again. Here we check the 792 * page tables (or emulated bits, if any), given the upper 793 * level VM system not knowing anything about existing 794 * references. 795 */ 796 } else if (((m->flags & PG_REFERENCED) == 0) && 797 (actcount = pmap_ts_referenced(m))) { 798 vm_page_activate(m); 799 VM_OBJECT_UNLOCK(object); 800 m->act_count += (actcount + ACT_ADVANCE); 801 continue; 802 } 803 804 /* 805 * If the upper level VM system knows about any page 806 * references, we activate the page. We also set the 807 * "activation count" higher than normal so that we will less 808 * likely place pages back onto the inactive queue again. 809 */ 810 if ((m->flags & PG_REFERENCED) != 0) { 811 vm_page_flag_clear(m, PG_REFERENCED); 812 actcount = pmap_ts_referenced(m); 813 vm_page_activate(m); 814 VM_OBJECT_UNLOCK(object); 815 m->act_count += (actcount + ACT_ADVANCE + 1); 816 continue; 817 } 818 819 /* 820 * If the upper level VM system does not believe that the page 821 * is fully dirty, but it is mapped for write access, then we 822 * consult the pmap to see if the page's dirty status should 823 * be updated. 824 */ 825 if (m->dirty != VM_PAGE_BITS_ALL && 826 (m->flags & PG_WRITEABLE) != 0) { 827 /* 828 * Avoid a race condition: Unless write access is 829 * removed from the page, another processor could 830 * modify it before all access is removed by the call 831 * to vm_page_cache() below. If vm_page_cache() finds 832 * that the page has been modified when it removes all 833 * access, it panics because it cannot cache dirty 834 * pages. In principle, we could eliminate just write 835 * access here rather than all access. In the expected 836 * case, when there are no last instant modifications 837 * to the page, removing all access will be cheaper 838 * overall. 839 */ 840 if (pmap_is_modified(m)) 841 vm_page_dirty(m); 842 else if (m->dirty == 0) 843 pmap_remove_all(m); 844 } 845 846 if (m->valid == 0) { 847 /* 848 * Invalid pages can be easily freed 849 */ 850 vm_page_free(m); 851 cnt.v_dfree++; 852 --page_shortage; 853 } else if (m->dirty == 0) { 854 /* 855 * Clean pages can be placed onto the cache queue. 856 * This effectively frees them. 857 */ 858 vm_page_cache(m); 859 --page_shortage; 860 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 861 /* 862 * Dirty pages need to be paged out, but flushing 863 * a page is extremely expensive verses freeing 864 * a clean page. Rather then artificially limiting 865 * the number of pages we can flush, we instead give 866 * dirty pages extra priority on the inactive queue 867 * by forcing them to be cycled through the queue 868 * twice before being flushed, after which the 869 * (now clean) page will cycle through once more 870 * before being freed. This significantly extends 871 * the thrash point for a heavily loaded machine. 872 */ 873 vm_page_flag_set(m, PG_WINATCFLS); 874 vm_page_requeue(m); 875 } else if (maxlaunder > 0) { 876 /* 877 * We always want to try to flush some dirty pages if 878 * we encounter them, to keep the system stable. 879 * Normally this number is small, but under extreme 880 * pressure where there are insufficient clean pages 881 * on the inactive queue, we may have to go all out. 882 */ 883 int swap_pageouts_ok, vfslocked = 0; 884 struct vnode *vp = NULL; 885 struct mount *mp = NULL; 886 887 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 888 swap_pageouts_ok = 1; 889 } else { 890 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 891 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 892 vm_page_count_min()); 893 894 } 895 896 /* 897 * We don't bother paging objects that are "dead". 898 * Those objects are in a "rundown" state. 899 */ 900 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 901 VM_OBJECT_UNLOCK(object); 902 vm_page_requeue(m); 903 continue; 904 } 905 906 /* 907 * Following operations may unlock 908 * vm_page_queue_mtx, invalidating the 'next' 909 * pointer. To prevent an inordinate number 910 * of restarts we use our marker to remember 911 * our place. 912 * 913 */ 914 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 915 m, &marker, pageq); 916 /* 917 * The object is already known NOT to be dead. It 918 * is possible for the vget() to block the whole 919 * pageout daemon, but the new low-memory handling 920 * code should prevent it. 921 * 922 * The previous code skipped locked vnodes and, worse, 923 * reordered pages in the queue. This results in 924 * completely non-deterministic operation and, on a 925 * busy system, can lead to extremely non-optimal 926 * pageouts. For example, it can cause clean pages 927 * to be freed and dirty pages to be moved to the end 928 * of the queue. Since dirty pages are also moved to 929 * the end of the queue once-cleaned, this gives 930 * way too large a weighting to defering the freeing 931 * of dirty pages. 932 * 933 * We can't wait forever for the vnode lock, we might 934 * deadlock due to a vn_read() getting stuck in 935 * vm_wait while holding this vnode. We skip the 936 * vnode if we can't get it in a reasonable amount 937 * of time. 938 */ 939 if (object->type == OBJT_VNODE) { 940 vp = object->handle; 941 if (vp->v_type == VREG && 942 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 943 mp = NULL; 944 ++pageout_lock_miss; 945 if (object->flags & OBJ_MIGHTBEDIRTY) 946 vnodes_skipped++; 947 goto unlock_and_continue; 948 } 949 KASSERT(mp != NULL, 950 ("vp %p with NULL v_mount", vp)); 951 vm_page_unlock_queues(); 952 vm_object_reference_locked(object); 953 VM_OBJECT_UNLOCK(object); 954 vfslocked = VFS_LOCK_GIANT(vp->v_mount); 955 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK, 956 curthread)) { 957 VM_OBJECT_LOCK(object); 958 vm_page_lock_queues(); 959 ++pageout_lock_miss; 960 if (object->flags & OBJ_MIGHTBEDIRTY) 961 vnodes_skipped++; 962 vp = NULL; 963 goto unlock_and_continue; 964 } 965 VM_OBJECT_LOCK(object); 966 vm_page_lock_queues(); 967 /* 968 * The page might have been moved to another 969 * queue during potential blocking in vget() 970 * above. The page might have been freed and 971 * reused for another vnode. 972 */ 973 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE || 974 m->object != object || 975 TAILQ_NEXT(m, pageq) != &marker) { 976 if (object->flags & OBJ_MIGHTBEDIRTY) 977 vnodes_skipped++; 978 goto unlock_and_continue; 979 } 980 981 /* 982 * The page may have been busied during the 983 * blocking in vget(). We don't move the 984 * page back onto the end of the queue so that 985 * statistics are more correct if we don't. 986 */ 987 if (m->busy || (m->oflags & VPO_BUSY)) { 988 goto unlock_and_continue; 989 } 990 991 /* 992 * If the page has become held it might 993 * be undergoing I/O, so skip it 994 */ 995 if (m->hold_count) { 996 vm_page_requeue(m); 997 if (object->flags & OBJ_MIGHTBEDIRTY) 998 vnodes_skipped++; 999 goto unlock_and_continue; 1000 } 1001 } 1002 1003 /* 1004 * If a page is dirty, then it is either being washed 1005 * (but not yet cleaned) or it is still in the 1006 * laundry. If it is still in the laundry, then we 1007 * start the cleaning operation. 1008 * 1009 * decrement page_shortage on success to account for 1010 * the (future) cleaned page. Otherwise we could wind 1011 * up laundering or cleaning too many pages. 1012 */ 1013 if (vm_pageout_clean(m) != 0) { 1014 --page_shortage; 1015 --maxlaunder; 1016 } 1017unlock_and_continue: 1018 VM_OBJECT_UNLOCK(object); 1019 if (mp != NULL) { 1020 vm_page_unlock_queues(); 1021 if (vp != NULL) 1022 vput(vp); 1023 VFS_UNLOCK_GIANT(vfslocked); 1024 vm_object_deallocate(object); 1025 vn_finished_write(mp); 1026 vm_page_lock_queues(); 1027 } 1028 next = TAILQ_NEXT(&marker, pageq); 1029 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1030 &marker, pageq); 1031 continue; 1032 } 1033 VM_OBJECT_UNLOCK(object); 1034 } 1035 1036 /* 1037 * Compute the number of pages we want to try to move from the 1038 * active queue to the inactive queue. 1039 */ 1040 page_shortage = vm_paging_target() + 1041 cnt.v_inactive_target - cnt.v_inactive_count; 1042 page_shortage += addl_page_shortage; 1043 1044 /* 1045 * Scan the active queue for things we can deactivate. We nominally 1046 * track the per-page activity counter and use it to locate 1047 * deactivation candidates. 1048 */ 1049 pcount = cnt.v_active_count; 1050 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1051 1052 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1053 1054 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1055 ("vm_pageout_scan: page %p isn't active", m)); 1056 1057 next = TAILQ_NEXT(m, pageq); 1058 object = m->object; 1059 if ((m->flags & PG_MARKER) != 0) { 1060 m = next; 1061 continue; 1062 } 1063 if (!VM_OBJECT_TRYLOCK(object) && 1064 !vm_pageout_fallback_object_lock(m, &next)) { 1065 VM_OBJECT_UNLOCK(object); 1066 m = next; 1067 continue; 1068 } 1069 1070 /* 1071 * Don't deactivate pages that are busy. 1072 */ 1073 if ((m->busy != 0) || 1074 (m->oflags & VPO_BUSY) || 1075 (m->hold_count != 0)) { 1076 VM_OBJECT_UNLOCK(object); 1077 vm_page_requeue(m); 1078 m = next; 1079 continue; 1080 } 1081 1082 /* 1083 * The count for pagedaemon pages is done after checking the 1084 * page for eligibility... 1085 */ 1086 cnt.v_pdpages++; 1087 1088 /* 1089 * Check to see "how much" the page has been used. 1090 */ 1091 actcount = 0; 1092 if (object->ref_count != 0) { 1093 if (m->flags & PG_REFERENCED) { 1094 actcount += 1; 1095 } 1096 actcount += pmap_ts_referenced(m); 1097 if (actcount) { 1098 m->act_count += ACT_ADVANCE + actcount; 1099 if (m->act_count > ACT_MAX) 1100 m->act_count = ACT_MAX; 1101 } 1102 } 1103 1104 /* 1105 * Since we have "tested" this bit, we need to clear it now. 1106 */ 1107 vm_page_flag_clear(m, PG_REFERENCED); 1108 1109 /* 1110 * Only if an object is currently being used, do we use the 1111 * page activation count stats. 1112 */ 1113 if (actcount && (object->ref_count != 0)) { 1114 vm_page_requeue(m); 1115 } else { 1116 m->act_count -= min(m->act_count, ACT_DECLINE); 1117 if (vm_pageout_algorithm || 1118 object->ref_count == 0 || 1119 m->act_count == 0) { 1120 page_shortage--; 1121 if (object->ref_count == 0) { 1122 KASSERT(!pmap_page_is_mapped(m), 1123 ("vm_pageout_scan: page %p is mapped", m)); 1124 if (m->dirty == 0) 1125 vm_page_cache(m); 1126 else 1127 vm_page_deactivate(m); 1128 } else { 1129 vm_page_deactivate(m); 1130 } 1131 } else { 1132 vm_page_requeue(m); 1133 } 1134 } 1135 VM_OBJECT_UNLOCK(object); 1136 m = next; 1137 } 1138 vm_page_unlock_queues(); 1139#if !defined(NO_SWAPPING) 1140 /* 1141 * Idle process swapout -- run once per second. 1142 */ 1143 if (vm_swap_idle_enabled) { 1144 static long lsec; 1145 if (time_second != lsec) { 1146 vm_req_vmdaemon(VM_SWAP_IDLE); 1147 lsec = time_second; 1148 } 1149 } 1150#endif 1151 1152 /* 1153 * If we didn't get enough free pages, and we have skipped a vnode 1154 * in a writeable object, wakeup the sync daemon. And kick swapout 1155 * if we did not get enough free pages. 1156 */ 1157 if (vm_paging_target() > 0) { 1158 if (vnodes_skipped && vm_page_count_min()) 1159 (void) speedup_syncer(); 1160#if !defined(NO_SWAPPING) 1161 if (vm_swap_enabled && vm_page_count_target()) 1162 vm_req_vmdaemon(VM_SWAP_NORMAL); 1163#endif 1164 } 1165 1166 /* 1167 * If we are critically low on one of RAM or swap and low on 1168 * the other, kill the largest process. However, we avoid 1169 * doing this on the first pass in order to give ourselves a 1170 * chance to flush out dirty vnode-backed pages and to allow 1171 * active pages to be moved to the inactive queue and reclaimed. 1172 */ 1173 if (pass != 0 && 1174 ((swap_pager_avail < 64 && vm_page_count_min()) || 1175 (swap_pager_full && vm_paging_target() > 0))) 1176 vm_pageout_oom(VM_OOM_MEM); 1177} 1178 1179 1180void 1181vm_pageout_oom(int shortage) 1182{ 1183 struct proc *p, *bigproc; 1184 vm_offset_t size, bigsize; 1185 struct thread *td; 1186 struct vmspace *vm; 1187 1188 /* 1189 * We keep the process bigproc locked once we find it to keep anyone 1190 * from messing with it; however, there is a possibility of 1191 * deadlock if process B is bigproc and one of it's child processes 1192 * attempts to propagate a signal to B while we are waiting for A's 1193 * lock while walking this list. To avoid this, we don't block on 1194 * the process lock but just skip a process if it is already locked. 1195 */ 1196 bigproc = NULL; 1197 bigsize = 0; 1198 sx_slock(&allproc_lock); 1199 FOREACH_PROC_IN_SYSTEM(p) { 1200 int breakout; 1201 1202 if (PROC_TRYLOCK(p) == 0) 1203 continue; 1204 /* 1205 * If this is a system, protected or killed process, skip it. 1206 */ 1207 if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) || 1208 (p->p_pid == 1) || P_KILLED(p) || 1209 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1210 PROC_UNLOCK(p); 1211 continue; 1212 } 1213 /* 1214 * If the process is in a non-running type state, 1215 * don't touch it. Check all the threads individually. 1216 */ 1217 breakout = 0; 1218 FOREACH_THREAD_IN_PROC(p, td) { 1219 thread_lock(td); 1220 if (!TD_ON_RUNQ(td) && 1221 !TD_IS_RUNNING(td) && 1222 !TD_IS_SLEEPING(td)) { 1223 thread_unlock(td); 1224 breakout = 1; 1225 break; 1226 } 1227 thread_unlock(td); 1228 } 1229 if (breakout) { 1230 PROC_UNLOCK(p); 1231 continue; 1232 } 1233 /* 1234 * get the process size 1235 */ 1236 vm = vmspace_acquire_ref(p); 1237 if (vm == NULL) { 1238 PROC_UNLOCK(p); 1239 continue; 1240 } 1241 if (!vm_map_trylock_read(&vm->vm_map)) { 1242 vmspace_free(vm); 1243 PROC_UNLOCK(p); 1244 continue; 1245 } 1246 size = vmspace_swap_count(vm); 1247 vm_map_unlock_read(&vm->vm_map); 1248 if (shortage == VM_OOM_MEM) 1249 size += vmspace_resident_count(vm); 1250 vmspace_free(vm); 1251 /* 1252 * if the this process is bigger than the biggest one 1253 * remember it. 1254 */ 1255 if (size > bigsize) { 1256 if (bigproc != NULL) 1257 PROC_UNLOCK(bigproc); 1258 bigproc = p; 1259 bigsize = size; 1260 } else 1261 PROC_UNLOCK(p); 1262 } 1263 sx_sunlock(&allproc_lock); 1264 if (bigproc != NULL) { 1265 killproc(bigproc, "out of swap space"); 1266 sched_nice(bigproc, PRIO_MIN); 1267 PROC_UNLOCK(bigproc); 1268 wakeup(&cnt.v_free_count); 1269 } 1270} 1271 1272/* 1273 * This routine tries to maintain the pseudo LRU active queue, 1274 * so that during long periods of time where there is no paging, 1275 * that some statistic accumulation still occurs. This code 1276 * helps the situation where paging just starts to occur. 1277 */ 1278static void 1279vm_pageout_page_stats() 1280{ 1281 vm_object_t object; 1282 vm_page_t m,next; 1283 int pcount,tpcount; /* Number of pages to check */ 1284 static int fullintervalcount = 0; 1285 int page_shortage; 1286 1287 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1288 page_shortage = 1289 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1290 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1291 1292 if (page_shortage <= 0) 1293 return; 1294 1295 pcount = cnt.v_active_count; 1296 fullintervalcount += vm_pageout_stats_interval; 1297 if (fullintervalcount < vm_pageout_full_stats_interval) { 1298 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count / 1299 cnt.v_page_count; 1300 if (pcount > tpcount) 1301 pcount = tpcount; 1302 } else { 1303 fullintervalcount = 0; 1304 } 1305 1306 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1307 while ((m != NULL) && (pcount-- > 0)) { 1308 int actcount; 1309 1310 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1311 ("vm_pageout_page_stats: page %p isn't active", m)); 1312 1313 next = TAILQ_NEXT(m, pageq); 1314 object = m->object; 1315 1316 if ((m->flags & PG_MARKER) != 0) { 1317 m = next; 1318 continue; 1319 } 1320 if (!VM_OBJECT_TRYLOCK(object) && 1321 !vm_pageout_fallback_object_lock(m, &next)) { 1322 VM_OBJECT_UNLOCK(object); 1323 m = next; 1324 continue; 1325 } 1326 1327 /* 1328 * Don't deactivate pages that are busy. 1329 */ 1330 if ((m->busy != 0) || 1331 (m->oflags & VPO_BUSY) || 1332 (m->hold_count != 0)) { 1333 VM_OBJECT_UNLOCK(object); 1334 vm_page_requeue(m); 1335 m = next; 1336 continue; 1337 } 1338 1339 actcount = 0; 1340 if (m->flags & PG_REFERENCED) { 1341 vm_page_flag_clear(m, PG_REFERENCED); 1342 actcount += 1; 1343 } 1344 1345 actcount += pmap_ts_referenced(m); 1346 if (actcount) { 1347 m->act_count += ACT_ADVANCE + actcount; 1348 if (m->act_count > ACT_MAX) 1349 m->act_count = ACT_MAX; 1350 vm_page_requeue(m); 1351 } else { 1352 if (m->act_count == 0) { 1353 /* 1354 * We turn off page access, so that we have 1355 * more accurate RSS stats. We don't do this 1356 * in the normal page deactivation when the 1357 * system is loaded VM wise, because the 1358 * cost of the large number of page protect 1359 * operations would be higher than the value 1360 * of doing the operation. 1361 */ 1362 pmap_remove_all(m); 1363 vm_page_deactivate(m); 1364 } else { 1365 m->act_count -= min(m->act_count, ACT_DECLINE); 1366 vm_page_requeue(m); 1367 } 1368 } 1369 VM_OBJECT_UNLOCK(object); 1370 m = next; 1371 } 1372} 1373 1374/* 1375 * vm_pageout is the high level pageout daemon. 1376 */ 1377static void 1378vm_pageout() 1379{ 1380 int error, pass; 1381 1382 /* 1383 * Initialize some paging parameters. 1384 */ 1385 cnt.v_interrupt_free_min = 2; 1386 if (cnt.v_page_count < 2000) 1387 vm_pageout_page_count = 8; 1388 1389 /* 1390 * v_free_reserved needs to include enough for the largest 1391 * swap pager structures plus enough for any pv_entry structs 1392 * when paging. 1393 */ 1394 if (cnt.v_page_count > 1024) 1395 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1396 else 1397 cnt.v_free_min = 4; 1398 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1399 cnt.v_interrupt_free_min; 1400 cnt.v_free_reserved = vm_pageout_page_count + 1401 cnt.v_pageout_free_min + (cnt.v_page_count / 768); 1402 cnt.v_free_severe = cnt.v_free_min / 2; 1403 cnt.v_free_min += cnt.v_free_reserved; 1404 cnt.v_free_severe += cnt.v_free_reserved; 1405 1406 /* 1407 * v_free_target and v_cache_min control pageout hysteresis. Note 1408 * that these are more a measure of the VM cache queue hysteresis 1409 * then the VM free queue. Specifically, v_free_target is the 1410 * high water mark (free+cache pages). 1411 * 1412 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1413 * low water mark, while v_free_min is the stop. v_cache_min must 1414 * be big enough to handle memory needs while the pageout daemon 1415 * is signalled and run to free more pages. 1416 */ 1417 if (cnt.v_free_count > 6144) 1418 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1419 else 1420 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1421 1422 if (cnt.v_free_count > 2048) { 1423 cnt.v_cache_min = cnt.v_free_target; 1424 cnt.v_cache_max = 2 * cnt.v_cache_min; 1425 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1426 } else { 1427 cnt.v_cache_min = 0; 1428 cnt.v_cache_max = 0; 1429 cnt.v_inactive_target = cnt.v_free_count / 4; 1430 } 1431 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1432 cnt.v_inactive_target = cnt.v_free_count / 3; 1433 1434 /* XXX does not really belong here */ 1435 if (vm_page_max_wired == 0) 1436 vm_page_max_wired = cnt.v_free_count / 3; 1437 1438 if (vm_pageout_stats_max == 0) 1439 vm_pageout_stats_max = cnt.v_free_target; 1440 1441 /* 1442 * Set interval in seconds for stats scan. 1443 */ 1444 if (vm_pageout_stats_interval == 0) 1445 vm_pageout_stats_interval = 5; 1446 if (vm_pageout_full_stats_interval == 0) 1447 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1448 1449 swap_pager_swap_init(); 1450 pass = 0; 1451 /* 1452 * The pageout daemon is never done, so loop forever. 1453 */ 1454 while (TRUE) { 1455 /* 1456 * If we have enough free memory, wakeup waiters. Do 1457 * not clear vm_pages_needed until we reach our target, 1458 * otherwise we may be woken up over and over again and 1459 * waste a lot of cpu. 1460 */ 1461 mtx_lock(&vm_page_queue_free_mtx); 1462 if (vm_pages_needed && !vm_page_count_min()) { 1463 if (!vm_paging_needed()) 1464 vm_pages_needed = 0; 1465 wakeup(&cnt.v_free_count); 1466 } 1467 if (vm_pages_needed) { 1468 /* 1469 * Still not done, take a second pass without waiting 1470 * (unlimited dirty cleaning), otherwise sleep a bit 1471 * and try again. 1472 */ 1473 ++pass; 1474 if (pass > 1) 1475 msleep(&vm_pages_needed, 1476 &vm_page_queue_free_mtx, PVM, "psleep", 1477 hz / 2); 1478 } else { 1479 /* 1480 * Good enough, sleep & handle stats. Prime the pass 1481 * for the next run. 1482 */ 1483 if (pass > 1) 1484 pass = 1; 1485 else 1486 pass = 0; 1487 error = msleep(&vm_pages_needed, 1488 &vm_page_queue_free_mtx, PVM, "psleep", 1489 vm_pageout_stats_interval * hz); 1490 if (error && !vm_pages_needed) { 1491 mtx_unlock(&vm_page_queue_free_mtx); 1492 pass = 0; 1493 vm_page_lock_queues(); 1494 vm_pageout_page_stats(); 1495 vm_page_unlock_queues(); 1496 continue; 1497 } 1498 } 1499 if (vm_pages_needed) 1500 cnt.v_pdwakeups++; 1501 mtx_unlock(&vm_page_queue_free_mtx); 1502 vm_pageout_scan(pass); 1503 } 1504} 1505 1506/* 1507 * Unless the free page queue lock is held by the caller, this function 1508 * should be regarded as advisory. Specifically, the caller should 1509 * not msleep() on &cnt.v_free_count following this function unless 1510 * the free page queue lock is held until the msleep() is performed. 1511 */ 1512void 1513pagedaemon_wakeup() 1514{ 1515 1516 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1517 vm_pages_needed = 1; 1518 wakeup(&vm_pages_needed); 1519 } 1520} 1521 1522#if !defined(NO_SWAPPING) 1523static void 1524vm_req_vmdaemon(int req) 1525{ 1526 static int lastrun = 0; 1527 1528 mtx_lock(&vm_daemon_mtx); 1529 vm_pageout_req_swapout |= req; 1530 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1531 wakeup(&vm_daemon_needed); 1532 lastrun = ticks; 1533 } 1534 mtx_unlock(&vm_daemon_mtx); 1535} 1536 1537static void 1538vm_daemon() 1539{ 1540 struct rlimit rsslim; 1541 struct proc *p; 1542 struct thread *td; 1543 struct vmspace *vm; 1544 int breakout, swapout_flags; 1545 1546 while (TRUE) { 1547 mtx_lock(&vm_daemon_mtx); 1548 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0); 1549 swapout_flags = vm_pageout_req_swapout; 1550 vm_pageout_req_swapout = 0; 1551 mtx_unlock(&vm_daemon_mtx); 1552 if (swapout_flags) 1553 swapout_procs(swapout_flags); 1554 1555 /* 1556 * scan the processes for exceeding their rlimits or if 1557 * process is swapped out -- deactivate pages 1558 */ 1559 sx_slock(&allproc_lock); 1560 FOREACH_PROC_IN_SYSTEM(p) { 1561 vm_pindex_t limit, size; 1562 1563 /* 1564 * if this is a system process or if we have already 1565 * looked at this process, skip it. 1566 */ 1567 PROC_LOCK(p); 1568 if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { 1569 PROC_UNLOCK(p); 1570 continue; 1571 } 1572 /* 1573 * if the process is in a non-running type state, 1574 * don't touch it. 1575 */ 1576 breakout = 0; 1577 FOREACH_THREAD_IN_PROC(p, td) { 1578 thread_lock(td); 1579 if (!TD_ON_RUNQ(td) && 1580 !TD_IS_RUNNING(td) && 1581 !TD_IS_SLEEPING(td)) { 1582 thread_unlock(td); 1583 breakout = 1; 1584 break; 1585 } 1586 thread_unlock(td); 1587 } 1588 if (breakout) { 1589 PROC_UNLOCK(p); 1590 continue; 1591 } 1592 /* 1593 * get a limit 1594 */ 1595 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1596 limit = OFF_TO_IDX( 1597 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1598 1599 /* 1600 * let processes that are swapped out really be 1601 * swapped out set the limit to nothing (will force a 1602 * swap-out.) 1603 */ 1604 if ((p->p_flag & P_INMEM) == 0) 1605 limit = 0; /* XXX */ 1606 vm = vmspace_acquire_ref(p); 1607 PROC_UNLOCK(p); 1608 if (vm == NULL) 1609 continue; 1610 1611 size = vmspace_resident_count(vm); 1612 if (limit >= 0 && size >= limit) { 1613 vm_pageout_map_deactivate_pages( 1614 &vm->vm_map, limit); 1615 } 1616 vmspace_free(vm); 1617 } 1618 sx_sunlock(&allproc_lock); 1619 } 1620} 1621#endif /* !defined(NO_SWAPPING) */ 1622